Current practice in the development and evaluation of spoken language dialogue systems.
Niels Ole Bernsen, Laila Dybkjaer, Ulrich Heid · 1999
In this paper a fast and effective algorithmnamed equal feature variance sum(EFVS) frame-synchronous searching is presented for state decoding. EFVS controls the state transition by using only the feature variance of the speech, instead of by using the state dwell distribution. The basic hypothesis of this new algorithmis the equality of feature variance sumin each state of the speech. Given the boundaries of the speech recognition unit (SRU), EFVS can generate the state sequence without dynamic searching. In a continuous speech word recognition system, this novel algorithm reduces the error rate by 36.8 % and speed up the systemby 65.6 % compared with the traditional state decoding methods.